Dynamics Analysis And Forecasting For Fire Time Series

Many time-varying factors have contributed to the fire accident, such as climate, population density and level of economic development; the observed fire time series is the combined result of these dynamic factors. The occurrence of single fire is random, however the account of fires in certain region and period appears regular to some extent. In this dissertation, the research object is the number of fires, the number of deaths, the number of injuries and economic losses per capita from1950to2008. According to the inherent variation of the fire time series, fire time series prediction method applied to small samples is selected and high-precision forecasting model is established.Based on support vector regression and Markov state transition, a new prediction model termed as Markov-support vector regression (MSVR) model is proposed to forecast the fire time series. In this proposed model, a SVR is adapted to build an optimal prediction model from a series of fire data and then the Markov state transition is used to reduce the residuals errors produced by the mentioned model. The results show that the result performance got from the MSVR model is better than that from the pure SVR model and RBF model.Since the actual fire time series with non-linear and non-stationary characteristics is not suitable to build a SVR forecasting model, in order to reduce the impact of non-stationary, the fire time series is decomposed into a series of intrinsic mode functions (IMF) in different scale space by using ensemble empirical mode decomposition (EEMD). To overcome the problem that support vector regression model needs to set the embedding dimension in advance, a new combined forecasting model is proposed for fire time series based on a combination of EEMD and multivariate phase-space reconstruction. Multivariate phase-space reconstruction is used to reconstruct the phase space of IMF, based on which the embedding dimension of forecasting model is estimated. The method proposed above is well performed in prediction accuracy.